AWS SageMaker AI adds serverless fine-tuning for NVIDIA Nemotron 3 models

AWS SageMaker AI introduced serverless model customization for NVIDIA’s Nemotron 3 family, enabling supervised fine-tuning (SFT), reinforcement learning with verifiable rewards (RLVR), and reinforcement learning from AI feedback (RLAIF) without provisioning or managing infrastructure.

The Nemotron 3 architecture combines Mamba-2 layers for linear-time sequence processing, Transformer attention layers for associative recall, and Latent Mixture-of-Experts layers that compress tokens before routing to specialized experts, activating only a fraction of total parameters per forward pass.

Nemotron 3 Nano (30B total parameters, 3B active) and Nemotron 3 Super (120B total parameters, 12B active) are supported, with the Nano variant optimized for high-volume, multi-agent workloads and the Super variant designed for complex reasoning and sustained multi-step agentic tasks.

Serverless customization in SageMaker AI handles infrastructure provisioning and training orchestration, allowing users to focus on data curation, business use cases, and evaluation while paying only for consumed resources.

The supported fine-tuning techniques include SFT for teaching new behaviors via labeled input-output pairs, RLVR for optimizing against verifiable objectives such as tool calling accuracy or code correctness, and RLAIF for iterative policy improvement guided by a separate AI model when human evaluation is costly or subjective.